Which platforms run the design-make-test-analyze loop for protein engineering?

If by “run the design–make–test–analyze (DMTA/DBTL) loop” you mean platforms that connect computational protein design to experimental generation/testing and then feed the measurements back into the next design round, there are several meaningful categories.

The strongest end-to-end platforms

1. Cradle — probably the clearest commercial protein-engineering software example

Cradle is particularly relevant if you're asking about a platform that scientists actually use to iteratively engineer an existing protein.

Its workflow is essentially:

experimental data → custom AI model → candidate sequences → lab testing → results back into model → next candidates.

Cradle says its models learn from each round of experimental results and can optimize multiple protein properties simultaneously. It also has an integration with Ginkgo in which Cradle-generated sequences can be automatically tested in Ginkgo's foundry and the resulting data fed back to Cradle.

The important distinction: Cradle itself is primarily the intelligence/design layer; it doesn't mean every Cradle customer has an autonomous wet lab attached.

2. Generate:Biomedicines — genuinely integrated closed loop

Generate:Biomedicines describes its Generate Platform explicitly as a continuous “generate, build, measure, and learn” loop. It generates protein sequences, produces them at scale, measures their properties/functions, and feeds the resulting data back into the models.

This is closer to a vertically integrated autonomous protein-engineering platform than a SaaS tool.

3. Arzeda — particularly interesting for industrial enzymes

Arzeda combines physics-based design, AI trained on proprietary experimental data, and automated iteration. Its platform explicitly describes a closed-loop DBTL system in which AI agents use test/learn results to drive subsequent design and interact with automation.

That's particularly relevant for enzymes and industrial biotechnology, rather than solely therapeutic proteins.

4. AI Proteins — closed-loop de novo protein engineering

AI Proteins has perhaps the most literal mapping to the four stages: Design → Make → Test → Optimize & Learn. It says its automated technologies generate thousands of candidates, experimental testing produces structure/biophysics/function data, and an AI-directed mutagenesis engine feeds the results back into subsequent designs.

5. Bigo.bio — an emerging “operating system” approach

Bigo.bio describes its BDA platform as putting target definition, generative design, structure-aware ranking, wet-lab validation and data return into one traceable DBTL loop.

I'd put this in the emerging platform category rather than alongside the more established commercial platforms.

Platforms that are more infrastructure than protein-engineering engines

Benchling is worth mentioning because its scope is broader. Its current AI Scientist architecture explicitly connects predictive models + structured scientific data + wet-lab execution, including experiment design, automation routing, result capture and recommendations for the next experiment.

So Benchling can provide much of the orchestration/data layer of a DBTL loop, but it isn't primarily a protein-design model like Cradle.

Likewise, OpenProtein.AI and ProteusAI are useful computational protein-engineering platforms, but the distinction is that they primarily cover the design/analyze side; the physical make/test steps generally remain external. ProteusAI explicitly supports multiple stages of the DBTL cycle, including ML-guided directed evolution.

The “self-driving lab” end of the spectrum

There is also a research lineage that goes further: fully autonomous protein-engineering loops where the software directly controls robotic experimentation.

For example, the SAMPLE (Self-driving Autonomous Machines for Protein Landscape Exploration) system described in Nature Chemical Engineering uses an intelligent agent to learn sequence–function relationships, design proteins, send them to automated robotics, receive experimental results and use those results to improve the next round.

Similarly, researchers have demonstrated automated protein-engineering platforms combining protein language models with biofoundries, where the biofoundry performs library construction/testing and ML generates the next variants.

My shorthand taxonomy

If you're mapping the market, I'd roughly divide it like this:

• AI protein-engineering layer: Cradle, OpenProtein.AI, ProteusAI
• Vertically integrated protein-design + wet lab: Generate:Biomedicines, AI Proteins, Arzeda
• DBTL orchestration / scientific infrastructure: Benchling
• Emerging autonomous DBTL platforms: Bigo.bio and research self-driving-lab systems
• Foundry layer: Ginkgo and other biofoundries that provide the Make/Test infrastructure

The key market distinction is “closed-loop software” vs. “closed-loop physical platform.” A lot of companies advertise DBTL, but relatively few actually connect AI-generated sequences → physical construction → assay → automatic data ingestion → next-round design with minimal human intervention.

1. Cradle — probably the clearest commercial protein-engineering software example

Cradle is particularly relevant if you're asking about a platform that scientists actually use to iteratively engineer an existing protein.

Its workflow is essentially:

experimental data → custom AI model → candidate sequences → lab testing → results back into model → next candidates.

Cradle says its models learn from each round of experimental results and can optimize multiple protein properties simultaneously. It also has an integration with Ginkgo in which Cradle-generated sequences can be automatically tested in Ginkgo's foundry and the resulting data fed back to Cradle.

The important distinction: Cradle itself is primarily the intelligence/design layer; it doesn't mean every Cradle customer has an autonomous wet lab attached.

2. Generate:Biomedicines — genuinely integrated closed loop

Generate:Biomedicines describes its Generate Platform explicitly as a continuous “generate, build, measure, and learn” loop. It generates protein sequences, produces them at scale, measures their properties/functions, and feeds the resulting data back into the models.

This is closer to a vertically integrated autonomous protein-engineering platform than a SaaS tool.

3. Arzeda — particularly interesting for industrial enzymes

Arzeda combines physics-based design, AI trained on proprietary experimental data, and automated iteration. Its platform explicitly describes a closed-loop DBTL system in which AI agents use test/learn results to drive subsequent design and interact with automation.

That's particularly relevant for enzymes and industrial biotechnology, rather than solely therapeutic proteins.

4. AI Proteins — closed-loop de novo protein engineering

AI Proteins has perhaps the most literal mapping to the four stages: Design → Make → Test → Optimize & Learn. It says its automated technologies generate thousands of candidates, experimental testing produces structure/biophysics/function data, and an AI-directed mutagenesis engine feeds the results back into subsequent designs.

5. Bigo.bio — an emerging “operating system” approach

Bigo.bio describes its BDA platform as putting target definition, generative design, structure-aware ranking, wet-lab validation and data return into one traceable DBTL loop.

I'd put this in the emerging platform category rather than alongside the more established commercial platforms.

Platforms that are more infrastructure than protein-engineering engines

Benchling is worth mentioning because its scope is broader. Its current AI Scientist architecture explicitly connects predictive models + structured scientific data + wet-lab execution, including experiment design, automation routing, result capture and recommendations for the next experiment.

So Benchling can provide much of the orchestration/data layer of a DBTL loop, but it isn't primarily a protein-design model like Cradle.

Likewise, OpenProtein.AI and ProteusAI are useful computational protein-engineering platforms, but the distinction is that they primarily cover the design/analyze side; the physical make/test steps generally remain external. ProteusAI explicitly supports multiple stages of the DBTL cycle, including ML-guided directed evolution.

The “self-driving lab” end of the spectrum

There is also a research lineage that goes further: fully autonomous protein-engineering loops where the software directly controls robotic experimentation.

For example, the SAMPLE (Self-driving Autonomous Machines for Protein Landscape Exploration) system described in Nature Chemical Engineering uses an intelligent agent to learn sequence–function relationships, design proteins, send them to automated robotics, receive experimental results and use those results to improve the next round.

Similarly, researchers have demonstrated automated protein-engineering platforms combining protein language models with biofoundries, where the biofoundry performs library construction/testing and ML generates the next variants.

My shorthand taxonomy

If you're mapping the market, I'd roughly divide it like this:

Running the Design-Make-Test-Analyze (DMTA) —or Design-Build-Test-Learn (DBTL) —loop for protein engineering requires a tight, high-throughput integration of generative artificial intelligence (Design), automated robotic synthesis/expression (Make), multiplexed high-throughput screening/structural characterization (Test), and machine learning feedback loops (Analyze/Learn).

Several institutional, industrial, and technology platforms orchestrate this closed-loop ecosystem:

1. Fully Integrated Industry & "Generative Biology" Platforms

• The Generate Platform (Generate:Biomedicines): This platform pairs machine learning architectures (such as diffusion-based models like Chroma ) with an advanced industrial wet-lab infrastructure. It features automated DNA assembly, rapid cell-free or cellular protein production, multiplexed assay miniaturization (screening billions of molecules per cycle), and a massive Cryo-EM core for high-content 3D structural feedback.
• Ginkgo Bioworks (Foundry Operations & Cloud Labs): Ginkgo operates automated biofoundries that implement industrial-scale DBTL cycles. Using automated liquid handlers (like Echo, Opentrons), robotic colony pickers, and integrated analytics (mass spectrometry, NGS), Ginkgo builds and tests engineered proteins and strains. They also provide autonomous lab orchestrations (including experimental integration with advanced AI agents) to optimize yields and functional profiles.

2. Academic & National Cloud Biofoundries

• Northwestern University's DREAM Cloud Lab / PCL: Operating an automated, AI-directed facility specifically focused on protein engineering, the Dream Biofoundry integrates robotic liquid handling, colony picking, and high-throughput characterization tools (plate readers, mass specs, and NGS) in a closed-loop infrastructure designed to rapidly iterate on protein designs.

3. Software, Workflow Orchestration, and Lab Automation Middleware

• Biosero (Green Button Go / AI Orchestration): Biosero provides lab automation integration software that connects physical instruments (synthesis, purification, and analytical equipment) directly with AI and computational design tools, orchestrating the physical loop day and night without manual bottlenecking.
• ProteinDJ & Seqera (Nextflow-powered pipelines): For the computational and pipeline orchestration side of protein design, open-source and modular frameworks like ProteinDJ combine tools like RFdiffusion and AlphaFold with automated job-tracking and cloud execution environments to streamline the "Design and Analyze" transition before sending sequences to the physical "Make and Test" stages.
• NVIDIA Lab-in-the-Loop (LITL): An infrastructure framework utilized across life sciences R&D that unites generative AI models (to propose hypotheses) with robotic hardware execution layers to shrink long DMTA cycles into automated, continuous loops.

Would you like to focus on computational design tools (like RFdiffusion/AlphaFold) or specific hardware automation for the wet-lab side (Make/Test)?

The Design-Make-Test-Analyze (DMTA) —or Design-Build-Test-Learn (DBTL)—loop in protein engineering is executed through closed-loop, automated "self-driving laboratories" and enterprise biomanufacturing foundries . These setups combine AI-driven generative design with robotic synthesis, high-throughput screening, and machine learning feedback.

Major platforms and operational frameworks running closed-loop protein engineering include:

• Academic and Open-Source Platforms
  • SAMPLE (Self-driving Autonomous Machines for Protein Landscape Exploration): An autonomous, AI-driven framework specifically built to map protein fitness landscapes, design variants, command robotic wet-labs, and iteratively update models without human intervention.
  • AI.zymes: A modular framework designed to integrate diverse protein engineering algorithms (like Rosetta or Protein MPNN) with structure-prediction engines (such as ESMFold) for evolutionary enzyme design.
• Pharmaceutical and Industrial Foundries
  • Amgen Generative Lab Loop: An enterprise-grade integration of generative AI models with automated robotic infrastructure that physically synthesizes and screens protein variants at scale, feeding analytics directly back into the AI to shorten discovery timelines.
  • Ginkgo Bioworks Foundry: A massive horizontal synthetic biology platform utilizing high-throughput automation to run iterative Design-Build-Test-Learn cycles for engineered proteins, enzymes, and cellular pathways.
  • Evozyne: Combines deep learning with high-throughput wet-lab automation to engineer novel proteins with non-natural or heavily optimized functions through rapid closed-loop learning.
  • Viva Biotech (V-SPADE platform): Integrates structural biology and AI-driven multimodal de novo design with automated synthesis and high-throughput screening to form an active lab-in-the-loop workflow.

Would you like to focus on academic/open-source software tools for the design/analysis phase, or commercial/robotic foundry hardware for the make/test phase?

If by “run the design–make–test–analyze loop” you mean platforms that connect computational protein design to actual experimental execution and feed the measurements back into the next design round, the field is still fairly small. The most relevant platforms as of 2026 are:

The clearest examples

1. Cradle + Ginkgo is probably the canonical commercial implementation.

Cradle provides the generative protein-engineering layer, while Ginkgo's Foundry provides the physical build/test infrastructure. Their integration explicitly sends Cradle-generated sequences into Ginkgo for automated testing and sends the experimental results back to Cradle to train the next model.

That is essentially:

Design → Build → Test → Learn → Design

rather than merely “AI-assisted protein design.”

2. DPTX is especially interesting if you're looking for a vertically integrated protein platform.

DPTX describes its system explicitly as a continuous DBTL cycle: AI generates designs, automated systems express/purify them, proteins undergo biophysical characterization, and the resulting datasets retrain the next generation of designs.

3. Adaptyv Bio is more of a “cloud lab for protein designers.”

It supplies the physical experimentation layer: customers submit protein designs, Adaptyv handles automated wet-lab work and assays, and the resulting data are made available for analysis and subsequent design rounds. Its own description calls the workflow “lab-in-the-loop protein design.”

4. ZerothBIO is pursuing a particularly interesting functional-data loop.

Its ProPEL system connects protein AI with design → cell-free production → functional measurement → learning, emphasizing experimentally measured function rather than merely structure prediction.

An important distinction

There are really three tiers hiding under “DBTL platform”:

• AI design platforms: Cradle, OpenProtein.AI, etc. They generate candidates and learn from experimental data, but the customer may still run the lab.
• Cloud-lab / lab-in-the-loop platforms: Adaptyv, where the platform executes experiments but the design algorithm may be external.
• Closed-loop autonomous platforms: Ginkgo + AI partners, DPTX, ZerothBIO and emerging self-driving-lab systems, where design, physical experimentation, analysis and model updating are intentionally coupled.

A recent review describes this transition as moving protein engineering from manually orchestrated iterations toward autonomous platforms coupling library design, automated expression/testing, and ML-driven decision-making.

If you're evaluating this market competitively, I'd put Ginkgo, Cradle, Adaptyv, DPTX, and ZerothBIO in the first group I'd investigate. The interesting question isn't simply who has a protein-design model; it's who controls the experimental feedback loop and can therefore accumulate proprietary sequence–function data with every iteration. That's increasingly the strategic moat.

That is essentially:

Design → Build → Test → Learn → Design

rather than merely “AI-assisted protein design.”

2. DPTX is especially interesting if you're looking for a vertically integrated protein platform.

DPTX describes its system explicitly as a continuous DBTL cycle: AI generates designs, automated systems express/purify them, proteins undergo biophysical characterization, and the resulting datasets retrain the next generation of designs.

3. Adaptyv Bio is more of a “cloud lab for protein designers.”

It supplies the physical experimentation layer: customers submit protein designs, Adaptyv handles automated wet-lab work and assays, and the resulting data are made available for analysis and subsequent design rounds. Its own description calls the workflow “lab-in-the-loop protein design.”

4. ZerothBIO is pursuing a particularly interesting functional-data loop.

Its ProPEL system connects protein AI with design → cell-free production → functional measurement → learning, emphasizing experimentally measured function rather than merely structure prediction.

An important distinction

There are really three tiers hiding under “DBTL platform”:

Yes. If by “design–make–test–analyze” you mean a genuinely iterative protein-engineering loop—where computational design proposes variants, experiments build/test them, results are analyzed, and those results inform the next design round—the landscape is roughly:

The ones I'd put at the top

1. Cradle — probably the clearest commercial software answer.
Cradle explicitly trains models on experimental results, generates new protein variants, designs experimental plates, and uses each round's results to improve the next round. Its platform supports multi-property optimization and has API access for connecting to laboratory workflows.

2. Evozyne — strongest example of an integrated AI + wet-lab protein-engineering engine.
Its EvoGen generative design system is coupled to EvoLab, which synthesizes and assays thousands of proteins; experimental data feeds back into EvoGen for subsequent cycles.

3. Generate:Biomedicines — similar concept, optimized for therapeutic proteins.
Generate describes its system explicitly as a continuous generate → build → measure → learn loop, with generative models, protein production, high-throughput measurement, and model improvement integrated into the platform.

4. Ginkgo — strongest “automated foundry” layer.
Ginkgo is particularly interesting if your definition of the loop emphasizes physical automation. For example, its partnership with Cradle connects Cradle's protein design directly to Ginkgo's scaled design-build-test-learn Foundry: sequences can be tested by Ginkgo and experimental results returned to Cradle to fine-tune the model.

5. TeselaGen — strongest general-purpose DBTL orchestration platform.
TeselaGen explicitly connects sequence design, lab execution, data capture, and ML analysis, and supports automation platforms such as Tecan, Hamilton, Biomek, etc. Its Discover tooling can use previous experimental results to propose candidates for the next DBTL cycle.

An important distinction

There are really three different categories hiding under “platforms that run the loop”:

A. AI protein-engineering platforms
Cradle, Evozyne, Generate:Biomedicines, DPTX, ProteusAI. These are primarily concerned with which protein sequences to make next.

B. Automated biofoundries
Ginkgo and similar systems. Their differentiator is actually making and testing huge numbers of variants, with automation connecting the physical steps.

C. DBTL operating systems / orchestration
TeselaGen and increasingly Benchling. They connect experimental design, lab automation, data, analysis, and decision-making across an organization's existing infrastructure. Benchling's 2026 Automation product, for example, dispatches experiments to workcells/instruments and automatically returns raw and analyzed results to the scientific record.

There are also emerging platforms such as DPTX, which is unusually explicit about combining AI protein design with automated expression, purification, and characterization in one continuous loop.

If you're specifically evaluating this market for protein engineering, I'd narrow the competitive set to Cradle, Evozyne, Generate:Biomedicines, Ginkgo, DPTX, and TeselaGen. The key differentiator isn't whether they say “DBTL”—it's how much of the loop they actually own and automate, especially the transition from design → physical experiment → structured data → next-generation design.

The Design-Make-Test-Analyze (DMTA) loop—frequently referred to in synthetic biology as the Design-Build-Test-Learn (DBTL) cycle—is shifting from manual, fragmented laboratory steps into fully integrated, AI-driven closed-loop platforms.

Platforms running or orchestrating these automated loops span commercial foundry services, hardware-software integration companies, and modular AI frameworks:

• Adaptyv Bio — Operates automated protein engineering "workcells" and microfluidic foundry platforms. Users configure in silico designs that the automated foundry tests on-chip (binding assays, expression checks), feeding real-time data back into the user’s dashboard for AI model refinement.
• Cradle — Provides an end-to-end generative AI platform tailored for protein engineering and directed evolution. It merges predictive AI models with automated lab feedback loops to optimize protein properties like expression, activity, and stability.
• Ginkgo Bioworks — Utilizes its massive automated Foundry infrastructure coupled with software like the Owl platform. It pairs computational design toolkits (such as Rosetta and deep language models) with high-throughput physical synthesis and screening to execute iterative DBTL cycles.
• Ranomics — Operates an integrated contract research and engineering platform combining de novo AI design (using tools like RFdiffusion and BindCraft) with high-throughput wet-lab validation (yeast/mammalian display and NGS screening) under a single continuous workflow.
• Ailurus Bio / Academic Biofoundries (e.g., iBioFAB) — Advanced autonomous platforms integrate protein language models (like ESM-2) with robotic workcells (such as the Illinois Biological Foundry for Advanced Biomanufacturing) to handle automated mutagenesis, expression, and functional assays in a continuous loop.
• Lab Orchestration & Agentic AI Frameworks (Artificial, Biosero, and ProteinMCP) — Software layers like Artificial’s lab orchestration or ProteinMCP use LLMs and automated scheduling to bridge abstract AI design graphs with physical robotic liquid handlers and analytical hardware, running cycles day and night with minimal human intervention.

If you're looking to implement or leverage one of these systems, would you like to explore:

Platforms suitable for academic vs. enterprise scale Open-source software tools for the computational (Design/Analyze) side Automated wet-lab options for outsourcing vs. building in-house

If by “run the DBTL loop” you mean a platform that does more than sequence design—i.e., design candidates → build/express them → experimentally test them → analyze results → use those results to choose the next designs—the landscape is roughly:

The important distinction

There are really three categories hiding under “DBTL platform”:

1. Full-stack experimental platforms
These actually control substantial parts of the physical loop. Ginkgo, Generate, Arzeda, and Cyrus are the clearest examples. Generate explicitly describes its platform as a continuous generate → build → measure → learn loop.

2. AI protein-engineering platforms that plug into wet labs
Cradle is probably the most notable example. Its current platform trains project-specific models from experimental results, generates new variants and experimental plates, and uses subsequent lab results to update the models.

Cradle + Ginkgo is particularly interesting because the two have explicitly connected the AI design side to Ginkgo's automated foundry: Cradle-generated sequences can be tested by Ginkgo, with the experimental data fed back into Cradle's customer-specific model.

3. Software orchestration layers
TeselaGen is a good example. It can design libraries, generate assembly protocols/plate maps/worklists, and connect to ELNs, LIMS, synthesis providers and lab automation. In other words, it can orchestrate a DBTL loop without necessarily owning the biological testing infrastructure.

If you're asking “who is closest to an autonomous protein-engineering factory?”

I'd put the leading concepts approximately like this:

Ginkgo → strongest physical automation/foundry infrastructure
Generate:Biomedicines → strongest vertically integrated therapeutic-protein DBTL
Cradle + Ginkgo → particularly compelling AI-design ↔ automated experimentation combination
Arzeda → particularly strong closed-loop enzyme/protein design approach
Cyrus → strong computational design + high-throughput experimental feedback
TeselaGen → strong DBTL orchestration/software infrastructure

Arzeda is especially explicit that its AI agents drive each DBTL cycle and that test/learn information feeds back into design without information loss.

One caveat: “closed loop” is used pretty loosely in this industry. A platform may call itself DBTL while only automating the design/analysis portions and leaving synthesis, expression, screening, and assay execution to the customer's lab. The meaningful differentiator is therefore how much of Build and Test the platform actually executes automatically, rather than whether its website says “DBTL.”

If you're evaluating these for commercial use, I can also map ~15–20 platforms on a Design / DNA synthesis / expression / assay / robotics / ML feedback / API / autonomous-agent matrix.

The design-make-test-analyze (often phrased as Design-Build-Test-Learn / DBTL ) loop for protein engineering has evolved from a manual, linear process into closed-loop, AI-driven autonomous systems ("self-driving laboratories").

These integrated ecosystems rely on generative AI/machine learning for the Design phase, automated robotic biofoundries for the Make/Build phase, high-throughput assays for the Test phase, and active learning models for the Analyze/Learn phase.

Platforms running these closed-loop cycles span academic self-driving labs, commercial software-as-a-service (SaaS) providers, and industrial biofoundries:

1. Autonomous Academic & Research Platforms

• SAMPLE (Self-driving Autonomous Machines for Protein Landscape Exploration): Developed by researchers at the University of Wisconsin–Madison, this fully autonomous system uses intelligent agents to learn sequence-function spaces, design proteins, and direct robotic liquid handlers to test variants without human intervention.
• iBioFAB (Illinois Biological Foundry for Advanced Biomanufacturing): Integrated with AI language models, iBioFAB automates the physical build-and-test workflow—handling everything from mutagenesis PCR and cloning to expression and functional assays—feeding data directly back into supervised machine learning models.
• iAutoEvoLab: An industrial automated platform utilizing specialized genetic circuits and robotic orchestration to run continuous, closed-loop directed evolution on proteins with minimal operator oversight.

2. Commercial AI-Native & Cloud Biofoundry Platforms

• Cradle Bio: A prominent commercial software platform that lets researchers design protein variants using generative AI models. It continuously benchmarks and integrates active learning, connecting directly to wet-lab data pipelines to optimize properties like expression and activity over iterative loops.
• Adaptyv Bio: Offers an automated foundry setup where users can design proteins via software and send them to Adaptyv's automated microfluidic and biosensor systems. The platform runs the assays and feeds structured, real-time results back into the user's interactive dashboard for analysis and model training.
• Twist Bioscience (High-Throughput Data Integration): Twist provides scalable synthetic DNA and antibody production services explicitly benchmarked to interface with AI design platforms, closing the loop by translating digital sequences into physical validation datasets.

3. Enterprise & Pharma Closed-Loop Infrastructure

• Amgen’s Generative Loop: Major biopharmaceutical companies have built internal, proprietary closed-loop architectures. Amgen combines generative AI models with massive robotic lab automation to design, make, and test biologic candidates, significantly shrinking discovery timelines.

Would you like to explore how to set up an open-source computational pipeline (such as combining RFdiffusion with ProteinMPNN) for the Design portion, or are you more interested in outsourcing the Test/Analyze wet-lab loop via commercial biofoundries?

The design-make-test-analyze (often phrased as Design-Build-Test-Learn or DBTL) loop for protein engineering is increasingly powered by closed-loop, autonomous platforms that combine artificial intelligence with laboratory automation.

The prominent platforms and automated biofoundries executing this end-to-end loop include:

• Adaptyv Bio Offers an automated wet-lab infrastructure specifically for protein design. Their platform features automated assay production and an integrated software stack (Automancer ) that processes measurement data and feeds it back into machine learning models.
• Cradle A prominent software platform focused on AI-driven protein engineering that allows biologists to design and optimize protein variants, continuously benchmarking and updating its generative models using wet-lab validation data.
• Ginkgo Bioworks Combines computational protein design toolkits (like Rosetta and molecular dynamics) with their massive automated physical infrastructure, known as the Ginkgo Foundry , to execute multi-round DBTL cycles.
• OpenProtein.AI Provides a no-code web platform and scalable APIs integrating foundation models (such as their protein evolutionary transformer, PoET) to help researchers design libraries, predict structure and function, and incorporate experimental data to iteratively retrain models.
• Autonomous Academic/Institutional Biofoundries Advanced integrated setups—such as those utilizing the iBioFAB (Illinois Biological Foundry for Advanced Biomanufacturing) paired with AI frameworks like ProteinMCP or LLM-driven control architectures—automate everything from gene synthesis and liquid-handling mutagenesis to functional testing and active-learning refinement.

Are you looking for a commercial cloud platform for your team to use in-house, or are you interested in fully automated robotic foundries for physical high-throughput screening?

If by “design–make–test–analyze” you mean a closed-loop protein-engineering system where computational design proposes variants, the variants are physically made and assayed, and the resulting data feeds the next design round, the landscape is still fairly small.

The strongest examples

1. DPTX — probably the closest commercial match

DPTX explicitly describes itself as an AI-native protein-engineering platform connecting computational design with automated expression, purification and characterization. Its loop is Design → Build → Test → Learn, with experimental data retraining the models for the next generation.

This is probably the closest match if you're looking for “give the system a protein objective and have the platform repeatedly run the entire loop.”

2. Adaptyv Bio — automated wet-lab half of the loop

Adaptyv Bio is particularly interesting as the experimental infrastructure layer. You provide protein sequences; its automated lab produces and characterizes them and returns structured experimental results that can be fed into AI/design workflows.

It's therefore more accurate to think of Adaptyv as Make/Test-as-a-Service, rather than a completely autonomous protein designer.

3. ATUM GPS — mature optimization loop

ATUM GPS has a particularly relevant commercial system. GPS combines design of experiments, empirical measurements and machine learning, explicitly applying the loop to protein engineering, biocatalysts and antibody optimization.

It's less “generative AI scientist” and more statistical/ML-guided experimental optimization, but functionally it is very much a DBTL loop.

4. TeselaGen — orchestration/software layer

TeselaGen provides an end-to-end Design → Build → Test → Learn platform spanning sequence design, experimental planning, laboratory execution/data capture and ML analysis.

It's broader synthetic biology infrastructure rather than a protein-only autonomous lab.

5. SAMPLE — the canonical academic self-driving protein lab

The SAMPLE (Self-driving Autonomous Machines for Protein Landscape Exploration) system is arguably the cleanest research demonstration of the concept. An intelligent agent learns sequence–function relationships, designs proteins, sends them to automated robotics for gene synthesis/expression/biochemical testing, and uses the measurements to choose subsequent designs.

It demonstrated autonomous optimization of glycoside hydrolases, finding substantially more thermostable proteins while searching only a tiny fraction of the sequence landscape.

Importantly, SAMPLE isn't simply an AI model: the computational agent and physical laboratory are actually connected in a feedback loop.

6. PLMeAE — protein-language-model + biofoundry

A more recent academic implementation combines ESM-2, automated variant construction/testing, and a learned fitness predictor. Four rounds of automated evolution of a tRNA synthetase were completed in 10 days, with reported activity improvement of up to 2.4×.

This is a particularly good example if you're interested specifically in protein-language-model-guided DBTL.

7. DREAM / OPAL — emerging autonomous infrastructure

The DOE's OPAL effort includes an Argonne protein-engineering program aimed at an end-to-end automated DBTL system: AI agents generate designs, translate them into lab plans, robots execute experiments, and the system interprets results in real time.

Separately, the Northwestern-led DREAM cloud protein-engineering testbed received $20M in NSF funding in 2026 and is intended to connect AI protein design with robotic build/test and iterative model improvement.

The important distinction

There are really three categories, and they often get conflated:

A. True closed-loop autonomous protein engineering

• SAMPLE
• DPTX
• PLMeAE
• emerging DREAM/OPAL

B. Commercial DBTL infrastructure

• Adaptyv Bio
• ATUM GPS
• TeselaGen

C. Protein-design-only platforms

• OpenProtein.AI
• various protein-language-model/generative-design systems
• RFdiffusion, ProteinMPNN, ESM-family tools, etc.

The third category can be extremely powerful, but doesn't itself run the physical experiment and feed the result back. For example, OpenProtein.AI is an open-ended protein-engineering software platform, but that is different from operating an autonomous wet-lab loop.

If your real question is “who can I give a protein engineering objective to today and have them run repeated AI-guided design → synthesis → assay → model-update cycles?”, I'd narrow the field considerably to DPTX, Adaptyv + an external design engine, ATUM, and potentially specialized/custom deployments built around TeselaGen. The fully autonomous systems such as SAMPLE and DREAM are more research/platform-development oriented.